2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
5 Stockfish is free software: you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation, either version 3 of the License, or
8 (at your option) any later version.
10 Stockfish is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with this program. If not, see <http://www.gnu.org/licenses/>.
19 #ifndef STOCKFISH_SIMD_H_INCLUDED
20 #define STOCKFISH_SIMD_H_INCLUDED
23 # include <immintrin.h>
25 #elif defined(USE_SSE41)
26 # include <smmintrin.h>
28 #elif defined(USE_SSSE3)
29 # include <tmmintrin.h>
31 #elif defined(USE_SSE2)
32 # include <emmintrin.h>
34 #elif defined(USE_MMX)
35 # include <mmintrin.h>
37 #elif defined(USE_NEON)
38 # include <arm_neon.h>
41 namespace Stockfish::Simd {
43 #if defined (USE_AVX512)
45 [[maybe_unused]] static int m512_hadd(__m512i sum, int bias) {
46 return _mm512_reduce_add_epi32(sum) + bias;
51 sum0 = [zmm0.i128[0], zmm0.i128[1], zmm0.i128[2], zmm0.i128[3]]
52 sum1 = [zmm1.i128[0], zmm1.i128[1], zmm1.i128[2], zmm1.i128[3]]
53 sum2 = [zmm2.i128[0], zmm2.i128[1], zmm2.i128[2], zmm2.i128[3]]
54 sum3 = [zmm3.i128[0], zmm3.i128[1], zmm3.i128[2], zmm3.i128[3]]
58 reduce_add_epi32(zmm0.i128[0]), reduce_add_epi32(zmm1.i128[0]), reduce_add_epi32(zmm2.i128[0]), reduce_add_epi32(zmm3.i128[0]),
59 reduce_add_epi32(zmm0.i128[1]), reduce_add_epi32(zmm1.i128[1]), reduce_add_epi32(zmm2.i128[1]), reduce_add_epi32(zmm3.i128[1]),
60 reduce_add_epi32(zmm0.i128[2]), reduce_add_epi32(zmm1.i128[2]), reduce_add_epi32(zmm2.i128[2]), reduce_add_epi32(zmm3.i128[2]),
61 reduce_add_epi32(zmm0.i128[3]), reduce_add_epi32(zmm1.i128[3]), reduce_add_epi32(zmm2.i128[3]), reduce_add_epi32(zmm3.i128[3])
64 [[maybe_unused]] static __m512i m512_hadd128x16_interleave(
65 __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3) {
67 __m512i sum01a = _mm512_unpacklo_epi32(sum0, sum1);
68 __m512i sum01b = _mm512_unpackhi_epi32(sum0, sum1);
70 __m512i sum23a = _mm512_unpacklo_epi32(sum2, sum3);
71 __m512i sum23b = _mm512_unpackhi_epi32(sum2, sum3);
73 __m512i sum01 = _mm512_add_epi32(sum01a, sum01b);
74 __m512i sum23 = _mm512_add_epi32(sum23a, sum23b);
76 __m512i sum0123a = _mm512_unpacklo_epi64(sum01, sum23);
77 __m512i sum0123b = _mm512_unpackhi_epi64(sum01, sum23);
79 return _mm512_add_epi32(sum0123a, sum0123b);
82 [[maybe_unused]] static void m512_add_dpbusd_epi32(
87 # if defined (USE_VNNI)
88 acc = _mm512_dpbusd_epi32(acc, a, b);
90 __m512i product0 = _mm512_maddubs_epi16(a, b);
91 product0 = _mm512_madd_epi16(product0, _mm512_set1_epi16(1));
92 acc = _mm512_add_epi32(acc, product0);
96 [[maybe_unused]] static void m512_add_dpbusd_epi32x2(
98 __m512i a0, __m512i b0,
99 __m512i a1, __m512i b1) {
101 # if defined (USE_VNNI)
102 acc = _mm512_dpbusd_epi32(acc, a0, b0);
103 acc = _mm512_dpbusd_epi32(acc, a1, b1);
105 __m512i product0 = _mm512_maddubs_epi16(a0, b0);
106 __m512i product1 = _mm512_maddubs_epi16(a1, b1);
107 product0 = _mm512_madd_epi16(product0, _mm512_set1_epi16(1));
108 product1 = _mm512_madd_epi16(product1, _mm512_set1_epi16(1));
109 acc = _mm512_add_epi32(acc, _mm512_add_epi32(product0, product1));
115 #if defined (USE_AVX2)
117 [[maybe_unused]] static int m256_hadd(__m256i sum, int bias) {
118 __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
119 sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
120 sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
121 return _mm_cvtsi128_si32(sum128) + bias;
124 [[maybe_unused]] static void m256_add_dpbusd_epi32(
129 # if defined (USE_VNNI)
130 acc = _mm256_dpbusd_epi32(acc, a, b);
132 __m256i product0 = _mm256_maddubs_epi16(a, b);
133 product0 = _mm256_madd_epi16(product0, _mm256_set1_epi16(1));
134 acc = _mm256_add_epi32(acc, product0);
138 [[maybe_unused]] static void m256_add_dpbusd_epi32x2(
140 __m256i a0, __m256i b0,
141 __m256i a1, __m256i b1) {
143 # if defined (USE_VNNI)
144 acc = _mm256_dpbusd_epi32(acc, a0, b0);
145 acc = _mm256_dpbusd_epi32(acc, a1, b1);
147 __m256i product0 = _mm256_maddubs_epi16(a0, b0);
148 __m256i product1 = _mm256_maddubs_epi16(a1, b1);
149 product0 = _mm256_madd_epi16(product0, _mm256_set1_epi16(1));
150 product1 = _mm256_madd_epi16(product1, _mm256_set1_epi16(1));
151 acc = _mm256_add_epi32(acc, _mm256_add_epi32(product0, product1));
157 #if defined (USE_SSSE3)
159 [[maybe_unused]] static int m128_hadd(__m128i sum, int bias) {
160 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
161 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
162 return _mm_cvtsi128_si32(sum) + bias;
165 [[maybe_unused]] static void m128_add_dpbusd_epi32(
170 __m128i product0 = _mm_maddubs_epi16(a, b);
171 product0 = _mm_madd_epi16(product0, _mm_set1_epi16(1));
172 acc = _mm_add_epi32(acc, product0);
175 [[maybe_unused]] static void m128_add_dpbusd_epi32x2(
177 __m128i a0, __m128i b0,
178 __m128i a1, __m128i b1) {
180 __m128i product0 = _mm_maddubs_epi16(a0, b0);
181 __m128i product1 = _mm_maddubs_epi16(a1, b1);
182 product0 = _mm_madd_epi16(product0, _mm_set1_epi16(1));
183 product1 = _mm_madd_epi16(product1, _mm_set1_epi16(1));
184 acc = _mm_add_epi32(acc, _mm_add_epi32(product0, product1));
189 #if defined (USE_NEON_DOTPROD)
191 [[maybe_unused]] static void dotprod_m128_add_dpbusd_epi32x2(
193 int8x16_t a0, int8x16_t b0,
194 int8x16_t a1, int8x16_t b1) {
196 acc = vdotq_s32(acc, a0, b0);
197 acc = vdotq_s32(acc, a1, b1);
200 [[maybe_unused]] static void dotprod_m128_add_dpbusd_epi32(
202 int8x16_t a, int8x16_t b) {
204 acc = vdotq_s32(acc, a, b);
208 #if defined (USE_NEON)
210 [[maybe_unused]] static int neon_m128_reduce_add_epi32(int32x4_t s) {
212 return vaddvq_s32(s);
214 return s[0] + s[1] + s[2] + s[3];
218 [[maybe_unused]] static int neon_m128_hadd(int32x4_t sum, int bias) {
219 return neon_m128_reduce_add_epi32(sum) + bias;
222 [[maybe_unused]] static void neon_m128_add_dpbusd_epi32x2(
224 int8x8_t a0, int8x8_t b0,
225 int8x8_t a1, int8x8_t b1) {
227 int16x8_t product = vmull_s8(a0, b0);
228 product = vmlal_s8(product, a1, b1);
229 acc = vpadalq_s16(acc, product);
234 [[maybe_unused]] static void neon_m128_add_dpbusd_epi32(
236 int8x16_t a, int8x16_t b) {
238 int16x8_t product0 = vmull_s8(vget_low_s8(a), vget_low_s8(b));
239 int16x8_t product1 = vmull_high_s8(a, b);
240 int16x8_t sum = vpaddq_s16(product0, product1);
241 acc = vpadalq_s16(acc, sum);
246 #endif // STOCKFISH_SIMD_H_INCLUDED